Leveraging Large Language Models for VNF Resource Forecasting

Jing Su, Suku Nair, Leo Popokh

Abstract: The evolution of the Network [entbl]Function Virtualization (NFV) paradigm has revolutionized the way network services are deployed, managed, and scaled. Within this transformative landscape, Virtual Network Function (VNF) resource prediction emerges as a cornerstone for optimizing network resource allocation and ensuring service reliability and efficiency. Traditional resource forecasting methods often struggle to adapt to the dynamic and non-linear nature of changes in resource consumption patterns in modern telecommunication networks. We address this challenge by leveraging the inherent pattern recognition and next-token prediction capabilities of Large Language Model (LLM) without requiring any domain-specific fine-tuning. Our study utilizes Llama2 as the foundation model to evaluate the performance against widely used probability-based models on a public VNF dataset that encompasses real-world resource consumption data of various VNFs for comparative analysis. Our findings suggest that LLM offers a highly effective alternative for VNF resource forecasting, demonstrating significant potential in enhancing network resource management.